A structural framework for instance-specific dynamic selection in classifier ensembles

Abstract

Dynamic selection strategies applied to classifier ensembles—whether at the feature level or at the ensemble member level—have been extensively investigated as mechanisms for enhancing predictive performance. Individually, these approaches have demonstrated consistent effectiveness across a range of classification scenarios. Building upon this foundation, the present study proposes an integrated framework in which dynamic selection mechanisms are employed not merely as auxiliary components, but as structural determinants of the ensemble itself. In this configuration, each instance is classified through a tailored subset of attributes and classifiers, thereby promoting instance-specific model adaptation. Although the simultaneous incorporation of multiple dynamic processes may introduce additional computational overhead, this research examines strategies designed to mitigate such costs. In particular, a decision criterion was developed to partition test instances into two distinct groups, such that only a subset of instances is subjected to the complete dynamic processing pipeline. This selective routing mechanism aims to preserve the benefits of structural adaptivity while constraining unnecessary computational expenditure. Preliminary experimental findings suggest that the proposed integrated approach produces notable gains in classification accuracy. Furthermore, the adoption of the decision criterion substantially reduces overall processing time, with no statistically significant degradation in predictive performance. Collectively, these results indicate that structural integration of dynamic selection techniques constitutes a viable and efficient advancement in ensemble learning methodologies.